Latest / ai morning by thehype
ai morning #99 — ai agents solve 49 of 50 unpublished physics problems
A new physics benchmark of unpublished problems still got solved almost perfectly by the best AI agent, for pocket change per problem. Pulling web search away barely changed the score. That points at something uncomfortable about what these models already know.
Marcus walks through the physics benchmark results, the detail that accuracy barely dropped without search access, and what it means for builders trusting any benchmark labeled unpublished or novel.
In this episode:
00:00 Intro
01:35 AI solves 49/50 physics problems — GPT-6-Astra hits 49 of 50 and Opus 5.5 46 of 50 on unpublished…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.